Pi Zero Vision

Low-latency local-HDMI vision apps for Raspberry Pi Zero 2 W with OV5647 camera.

Goals

• Fullscreen local HDMI camera display
• Lowest practical latency
• Classical tracking baseline that actually works on Pi Zero 2 W
• Experimental ML backends for person and hand detection

Hardware

• Raspberry Pi Zero 2 W
• OV5647 CSI camera
• Local HDMI display

Proven conclusions

• SDL2 + KMSDRM + rpicam-vid is the correct display pipeline
• Classical tracking is the practical production path
• TFLite hand landmark inference is too slow on this board
• ONNX quantized person model had compatibility issues in this environment
• HOG is lightweight but unreliable for this camera/framing setup

Apps

• apps/v1_camera_viewer: working fullscreen camera viewer
• apps/v2_swipe_simple: basic motion swipe tracker
• apps/v3_swipe_blob: improved low-latency swipe tracker
• apps/v4_tflite_hand_experimental: experimental hand landmark ML
• apps/person_tracker_hog: HOG person detector
• apps/person_tracker_onnx: experimental ONNX person detector
• apps/v5_person_uhd: UHD model-based person detector with HDMI overlay confidence

Quick start

./scripts/install_base.sh
./scripts/run_v1.sh

Recommended production path

• Gesture tracking on Pi Zero 2 W: apps/v3_swipe_blob
• Person detection on Pi Zero 2 W: apps/v5_person_uhd (UHD ONNX model, confidence overlay)